import time
import sys
import math
import numpy as np
import pickle
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
def the_ratio(n, k):
m = 1
for i in range(1, n):
m = m * (k + i)
m = round(m / math.factorial(n - 1))
return m
def genset_k_do_increment(k, X):
a = np.shape(X)
lastNZ = np.amax(np.matmul((X>0), np.diag(range(1, k+1))), axis=1)
b = np.shape(lastNZ)
k1lastNZ = k+1-lastNZ
Y = np.repeat(X,k1lastNZ, axis=0)
c = np.shape(Y)
idxY = 1
for q in range(1,a[0]+1):
cs = lastNZ[q-1]
for c in range(cs,k+1):
Y[idxY+c-cs-1,c-1] += 1
idxY = idxY + k+1-cs
return Y
def genset_k_by_inc(n, k):
m = the_ratio(n, k)
X = np.zeros((m, n), dtype=np.int8, order='C')
for i in range(0, n):
X[i][i] = 1
sm = 1
mX = the_ratio(n, sm)
while (sm < k):
tm = time.time()
sm = sm + 1
mX1 = the_ratio(n, sm)
X[0:mX1-1, :] = genset_k_do_increment(n, X[0:mX - 1, :])
mX = mX1
print("iteration:%i -- %.2f [s]" % (k - sm, time.time() - tm))
X[-1,-1] = X[0,0]
return X
def generate_dataset(n, k):
start_time = time.time()
print("Generating simplex data",end='')
print('\n')
X = genset_k_by_inc(n, k)
print(" -- Done")
mn = np.shape(X)
print(np.sum(X[0, :]))
print(mn)
print("GenSimpBIN: %.2f [s]" % (time.time() - start_time))
return X
def save_bin_dataset(X, fname):
start_time = time.time()
print("Saving BIN file: %s" % fname, end='')
with open(fname, 'wb') as f:
pickle.dump(X, f)
print("SaveBIN: %.2f [s]" % (time.time() - start_time))
print(" -- Done")
def save_txt_dataset(X, fname):
start_time = time.time()
print("Saving TXT file: %s" % fname, end='')
np.savetxt(fname, X, fmt='%i')
print(" -- Done")
print("SaveTXT: %.2f [s]" % (time.time() - start_time))
def load_bin_dataset(fname):
print("Loading BIN file: %s" % fname, end='')
with open(fname, 'rb') as f:
X = pickle.load(f)
print(" -- Done")
mn = np.shape(X)
print(np.sum(X[0, :]))
print(mn)
return X
def load_txt_dataset(fname):
print("Loading TXT file: %s" % fname, end='')
X = np.loadtxt(fname, dtype=np.int8)
print(" -- Done")
mn = np.shape(X)
print(np.sum(X[0, :]))
print(mn)
return X
def get_group_ratio(df):
df['gr'] = (df.f_tp + df.f_fp + df.f_tn + df.f_fn) / (df.m_tp + df.m_fp + df.m_tn + df.m_fn + df.f_tp + df.f_fp + df.f_tn + df.f_fn)
return df
def get_imbalance_ratio(df):
df['ir'] = (df.f_tp + df.f_fn + df.m_tp + df.m_fn) / (df.f_tn + df.f_fn + df.m_tn + df.m_fn + df.f_fp +df.f_tp + df.m_fp + df.m_tp)
#IR DLA PRAWDZ KLAS
return df
def get_men_true_pos_rate(df):
df['m_tpr'] = df.m_tp/(df.m_tp + df.m_fn)
return df
def get_females_true_pos_rate(df):
df['f_tpr'] = df.f_tp/(df.f_tp + df.f_fn)
return df
#ratio of true positive rates
#equal opportunity
def get_equal_opp_ratio(df):
df['equal_opp_ratio'] = (df.f_tp/(df.f_tp + df.f_fn)) / (df.m_tp/(df.m_tp + df.m_fn))
return df
#equal opportunity
#difference of true positive rates
def get_equal_opp_diff(df):
df['equal_opp_diff'] = df.f_tpr - df.m_tpr
return df
#each group has the same probability of being classified with positive outcome
#difference of probabilities
def get_statistical_parity(df):
df["stat_parity"] = (df.f_tp + df.f_fp)/(df.f_tp + df.f_fp + df.f_tn + df.f_fn) - (df.m_tp + df.m_fp)/(df.m_tp + df.m_fp + df.m_tn + df.m_fn)
return df
#similiar to statistical parity, but using ratio
def get_disparate_impact(df):
df["disp_impact"] = ((df.f_tp + df.f_fp)/(df.f_tp + df.f_fp + df.f_tn + df.f_fn)) / ((df.m_tp + df.m_fp)/(df.m_tp + df.m_fp + df.m_tn + df.m_fn))
return df
#accuracy equality ratio
def get_acc_equality_ratio(df):
df["acc_equality_ratio"] = ((df.f_tp + df.f_tn)/(df.f_tp + df.f_fp + df.f_tn + df.f_fn)) / ((df.m_tp + df.m_tn)/(df.m_tp + df.m_fp + df.m_tn + df.m_fn))
return df
#accuracy equality difference
def get_acc_equality_diff(df):
df["acc_equality_diff"] = ((df.f_tp + df.f_tn)/(df.f_tp + df.f_fp + df.f_tn + df.f_fn)) - ((df.m_tp + df.m_tn)/(df.m_tp + df.m_fp + df.m_tn + df.m_fn))
return df
#predictive equality ratio
def get_pred_equality_ratio(df):
df["pred_equality_ratio"] = (df.f_fp/(df.f_fp + df.f_tn)) / (df.m_fp/(df.m_fp + df.m_tn))
return df
#predictive equality difference
def get_pred_equality_diff(df):
df["pred_equality_diff"] = (df.f_fp/(df.f_fp + df.f_tn)) - (df.m_fp/(df.m_fp + df.m_tn))
return df
#predictive parity ratio
def get_pred_parity_ratio(df):
df["pred_parity_ratio"] = (df.f_tp/(df.f_tp + df.f_fp)) / (df.m_tp/(df.m_tp + df.m_fp))
return df
#predictive parity difference
def get_pred_parity_diff(df):
df["pred_parity_diff"] = (df.f_tp/(df.f_tp + df.f_fp)) - (df.m_tp/(df.m_tp + df.m_fp))
return df
def create_heatmap(df, fair_measure):
plt.clf()
li = sns.color_palette("light:b", as_cmap=True)
df1 = df.groupby(['ir', fair_measure]).size().reset_index(name='counts')
ir_counts = df1.groupby(['ir'])['counts'].sum().reset_index(name='ir_counts')
df1 = pd.merge(df1, ir_counts, on='ir', how='left')
df1['probability'] = df1.counts / df1.ir_counts
heatm = sns.histplot(df1, x="ir", y=fair_measure, weights="probability", binwidth=(0.0625, 0.05), cmap=li)
norm = plt.Normalize(df1['probability'].min(), df1['probability'].max())
sm = plt.cm.ScalarMappable(cmap=li, norm=norm)
sm.set_array([])
heatm.figure.colorbar(sm)
fm_name = fm_full_names[fair_measure]
plt.title(str(fm_name))
plt.ylabel(f"Fairness Measure - {fm_name}")
plt.xlabel("Minority Ratio")
fig = heatm.get_figure()
fig.savefig(f"plots/heatmap_{fair_measure}.png")
return
def create_histogram_ir(df, ir_selected, fair_measure, bins_selected):
plt.clf()
df1 = df[df["ir"] == ir_selected]
df1 = df1[[fair_measure]]
hist = df1.hist(bins=bins_selected)
fm_name = fm_full_names[fair_measure]
plt.title(str(fm_name) + ' for ir = ' + str(round(ir_selected,2)))
plt.xlabel(f"Fairness Measure - {fm_name}")
plt.ylabel("Counts")
fig = hist[0][0].get_figure()
fig.savefig(f"plots/histogram_{fair_measure}_ir_{round(ir_selected,2)}.png")
return
def create_histogram_gr(df, gr_selected, fair_measure, bins_selected):
plt.clf()
df1 = df[df["gr"] == gr_selected]
df1 = df1[[fair_measure]]
hist = df1.hist(bins=bins_selected)
fm_name = fm_full_names[fair_measure]
plt.title(str(fm_name) + ' for gr = ' + str(round(gr_selected,2)))
plt.xlabel(f"Fairness Measure - {fm_name}")
plt.ylabel("Counts")
fig = hist[0][0].get_figure()
fig.savefig(f"plots/histogram_{fair_measure}_gr_{round(gr_selected,2)}.png")
return
def create_dataframe(nparray, k):
df = pd.DataFrame(nparray, columns=['m_tp', 'm_fp', 'm_tn', 'm_fn', 'f_tp', 'f_fp', 'f_tn', 'f_fn'])
get_group_ratio(df)
get_imbalance_ratio(df)
df.replace([np.inf, -np.inf], 0, inplace=True)
get_men_true_pos_rate(df)
get_females_true_pos_rate(df)
df.replace([np.inf, -np.inf], 0, inplace=True)
# calculate fairness measures
get_equal_opp_ratio(df)
get_equal_opp_diff(df)
get_acc_equality_diff(df)
get_acc_equality_ratio(df)
get_statistical_parity(df)
get_disparate_impact(df)
get_pred_equality_ratio(df)
get_pred_equality_diff(df)
get_pred_parity_ratio(df)
get_pred_parity_diff(df)
df.replace([np.inf, -np.inf], 0, inplace=True)
df.replace(np.NaN, 0, inplace=True)
return df
